using --rect to train, mAP a little bit low

[wzf19947]

Search before asking

• I have searched the YOLOv5 issues and discussions and found no similar questions.

Question

I trained with two models, the only difference is using --rect param or not , it turns out using --rect to train, the mAP is much lower and grow up lower than don't use it, is is normal?

Additional

No response

[UltralyticsAssistant]

👋 Hello @wzf19947, thank you for your interest in YOLOv5 🚀! I see you've encountered an issue with using the --rect parameter during training. This is an automated response to acknowledge your issue, and an Ultralytics engineer will assist you soon.

If this is a 🐛 Bug Report, please provide a minimum reproducible example to help us understand and debug it more effectively.

For questions related to custom training or parameter tuning, please provide as much detail as possible, such as dataset examples, training logs, and ensure you're following recommended practices to achieve the best training results.

Requirements

Ensure you have Python>=3.8.0 with all dependencies installed, including PyTorch>=1.8. You can do this by cloning the YOLOv5 repository and installing the necessary requirements using pip install -r requirements.txt.

Environments

YOLOv5 offers various verified environments such as Notebooks with free GPU access, Google Cloud's Deep Learning VM, Amazon's Deep Learning AMI, and Docker images, each with all dependencies, including CUDA, CUDNN, Python, and PyTorch, preinstalled.

Status

Check our Continuous Integration (CI) tests on GitHub Actions, which verify the correct operation of YOLOv5's training, validation, inference, export, and benchmarks on multiple platforms. If these tests are passing, YOLOv5 should operate correctly in its supported environments.

Feel free to provide the additional information to assist us in resolving your query. 😊

[pderrenger]

@wzf19947 using the --rect parameter can result in lower mAP because it alters the aspect ratio of images during training, which might affect the model's learning efficiency. It's recommended to use rectangular training only when dealing with datasets where maintaining the original aspect ratio is crucial, such as when objects have a consistent orientation or size. Otherwise, consider training without --rect for potentially better mAP results. For more insights, refer to the YOLOv5 documentation.